ResTreeNet: A Height-Aware LiDAR Tree Classification Model With Explainable AI for Forestry Applications

Abstract

Tree species classification plays a crucial role in forest management, biodiversity conservation, and ecological monitoring. Light detection and ranging (LiDAR) technology, capturing highly detailed 3D structural information of vegetation, has become a powerful tool for automated tree classification. Among LiDAR techniques, terrestrial LiDAR provides high-resolution point-cloud data by scanning trees from the ground level, enabling precise species identification. However, applying deep learning models to LiDAR-based tree classification remains challenging due to the computational complexity of existing 3D architectures, which often struggle with scalability and practical large-scale implementation. To address these critical limitations, we propose ResTreeNet, an efficient and lightweight deep learning model designed explicitly for tree classification using terrestrial LiDAR point clouds. Our innovative approach combines residual networks for hierarchical feature extraction, a height-based grouping strategy to enhance structural representation, and a parameterized geometric transformation module to improve translation invariance and model adaptability. This work integrates explainable artificial intelligence (XAI) techniques, including gradient-weighted class action mapping (Grad-CAM) visualizations, to provide transparent and interpretable insight into the classification reasoning of the model, addressing the critical need for understanding automated decision-making processes. The comprehensive evaluation on a terrestrial LiDAR dataset demonstrates the superior performance of ResTreeNet, achieving a state-of-the-art accuracy of 94.02% on samples with 1024-points, surpassing the existing models by 2.03%. The robust capabilities of the model are further validated by outstanding classification metrics, including precision (94.24%), recall (93.63%), and the F1-score (93.54%), ensuring a balanced and reliable approach to tree species classification. With its lightweight architecture (requiring only 0.47 million parameters) and computational efficiency, ResTreeNet is a practical solution for large-scale ecological research, offering an innovative approach to automated forest monitoring and sustainable resource management.